DeepSeek’s new AI model talks less, so a coding task costs 20 cents

DeepSeek‘s newest AI model finishes a typical coding task for about 20 cents, according to the cost chart the Chinese lab published alongside it. On that same chart, Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash land somewhere between $2 and more than $10 for each task. The update, called V4.1-Flash-1017, gets there mostly by saying less: DeepSeek trained it to answer with shorter outputs, and every word a model does not write is a word nobody pays for.

That matters to anyone who lets an AI agent loose on a codebase and pays by the token. An agent that tests its own code, fails, fixes it and tries again can burn through hundreds of answers on a single job, so a model that is ten times cheaper per task turns an expensive habit into a small line on the bill. On DeepSWE, the software-engineering benchmark behind that chart, the update scores 77.0%, ahead of the 74.0% DeepSeek lists for Opus 5 and the 73.0% it gives GPT-5.6 Sol. The previous V4.1 Flash scored 74.2% and cost around 50 cents a task.

The gain comes from how DeepSeek trained the model after its main training run. The method, which the company calls reinforcement learning with self-feedback, lets the model grade its own work using its internal confidence and self-written critiques instead of human raters or a separate judging model. DeepSeek says this makes it better at building whole projects from an empty folder, and its release notes show a 3D model of Beijing’s Forbidden City and a recreation of the puzzle game Portal 2 built in the Godot engine. Its post-training lead, Wu Yu, called it “the era of democratized wish-driven coding,” as Wccftech reported.

Unlike Claude or GPT, the model itself is public. DeepSeek put the weights on Hugging Face and on Alibaba’s ModelScope under the MIT license, which lets any company run it on its own servers and keep its code away from DeepSeek’s. That is not the same as running it on a laptop. The model has 552 billion parameters, of which it switches on only 8 billion per word while reading and 16 billion while writing, and the files take up around 510GB, twice the storage of a 256GB base-model laptop. Wccftech reported that the earlier V4.1 Flash hit a peak of 494 tokens per second on a home rig built from four Nvidia DGX Spark desktop AI machines, hardware few individuals own.

The headline numbers come from DeepSeek, and no independent lab has tested the update yet, as The Decoder pointed out. DeepSeek’s own table also shows where the cheap model still trails: Opus 5 scores 51.8 to its 42.3 on Terminal-Bench 4.0, 75.3 to 68.2 on NL2Repo, which asks a model to write a whole code repository, and 56.3 to 42.6 on the expert-level quiz Humanity’s Last Exam. A score on a benchmark also says little about whether the code is pleasant to maintain. And anyone using DeepSeek’s app or hosted API, rather than the downloaded weights, is sending prompts to a company based in China, which several governments have already barred from official devices.

Through the API, the model costs $0.30 per million tokens of fresh input and $1.20 per million tokens of output at peak times, and half that off-peak. Peak means 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so a European working morning pays full price while weekends are always discounted. Developers still calling the old deepseek-v4-flash name are already being served the new model without changing anything.

DeepSeek released V4.1-Flash-1017 on October 17 on its website, app and API. It also opened its in-house Search and Fetch APIs to all developers at $0.75 per 1,000 requests; calls made through DeepSeek Harness, Claude Code and Codex stay free until billing starts on October 31 at 16:00 UTC.

The next question is whether the following DeepSeek release will be as easy to download. A day before this one, China’s commerce and science ministries published draft export rules that would put model weights under control once a system passes 2 trillion parameters, 10^26 operations of training compute, or a place in the global top ten. Open-weight releases would get automatic approval on registration, according to The Decoder, and anything already public is exempt, so this model is out of reach of the rules. A model that beats Opus 5 on its own charts is exactly the kind Beijing would now have to count.

Tags: Hugging Face, DeepSeek, Claude Opus 5, AI coding, DeepSeek V4.1 Flash, open-weight AI